Environment
- Codex: v0.133.0
- OS: macOS 26.5 (25F71)
- Model: auto (9router)
Problem
Codex unconditionally loads ALL skills from multiple directories with a hard budget limit of 5440 chars (~2% context window). No config option to exclude or disable unwanted skills. When skill count exceeds ~20, mass description truncation occurs.
Evidence
From ~/.codex/log/codex-tui.log:
truncated skill metadata to fit skills context budget
budget_limit=5440
total_skills=119
included_skills=119
omitted_skills=0
truncated_description_chars_per_skill=232
truncated_skill_descriptions=103
Root Cause
- Unconditional multi-directory scan — Codex scans
~/.codex/skills/, ~/.codex/skills/.system/, and ~/.agents/skills/ with no filter
- No exclude/ignore config —
config.toml has no [skills] section, no mechanism to disable unwanted skills
.system/ auto-restore — imagegen, plugin-creator, skill-creator, skill-installer, openai-docs auto-restore on every launch, cannot be permanently removed
- Budget too small — 119 skills × ~232 chars = 103/119 descriptions truncated
Expected
skills.exclude = [...] or skills.include = [...] in config.toml
- Per-skill
enabled = false override
- Option to disable
.system/ skill auto-restore
- OR dynamic budget that scales with context window size
Impact
Truncated descriptions make skills undiscoverable — agent cannot understand what most skills actually do. Users with many skills (agent frameworks, integrations) are hardest hit.
Environment
Problem
Codex unconditionally loads ALL skills from multiple directories with a hard budget limit of 5440 chars (~2% context window). No config option to exclude or disable unwanted skills. When skill count exceeds ~20, mass description truncation occurs.
Evidence
From
~/.codex/log/codex-tui.log:Root Cause
~/.codex/skills/,~/.codex/skills/.system/, and~/.agents/skills/with no filterconfig.tomlhas no[skills]section, no mechanism to disable unwanted skills.system/auto-restore — imagegen, plugin-creator, skill-creator, skill-installer, openai-docs auto-restore on every launch, cannot be permanently removedExpected
skills.exclude = [...]orskills.include = [...]in config.tomlenabled = falseoverride.system/skill auto-restoreImpact
Truncated descriptions make skills undiscoverable — agent cannot understand what most skills actually do. Users with many skills (agent frameworks, integrations) are hardest hit.